The disclosure provides a feature compression algorithm based on neural network, including the following steps: S1, image data preparation: collecting facial images, and uniformly performing map processing to the facial images collected; S2, feature data acquisition: delivering the facial images processed into a face recognition system for face detection and feature extraction, and saving facial feature data; S3, setting up a neural network model; S4, model iteration training; S5, storing a parameter model; and S6, feature compression. The feature compression algorithm based on neural network of the disclosure can not only achieve compression of original feature data, but also retain its original semantic feature, which belongs to a higher-dimensional feature abstraction. The compressed feature data can be directly used.
Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A computer program product comprising a non-transitory computer readable medium encoded with an information processing program for use in an information processing device, the program when executed performs the following steps: S 1 , image data preparation: collecting facial images, and uniformly performing map processing to the facial images collected; S 2 , feature data acquisition: delivering the facial images processed into a face recognition system for face detection and feature extraction, and saving facial feature data; S 3 , setting up a neural network model for inputting the facial feature data; S 4 , model iteration training: adopting an own coding training method; S 5 , saving a parameter model of the neural network model; and S 6 , feature compression: obtaining a compressed feature through the saved parameter model, wherein the own coding training method particularly comprises the following steps: S 401 , compressing information x of an input layer to a hidden layer to obtain h; S 402 , generating x′ using the h of the hidden layer; S 403 , comparing the generated x′ with the x of the input layer to obtain a prediction error; S 404 , performing reverse delivery of the prediction error, and gradually improving an accuracy of feature compression; and S 405 , obtaining data of the h in the hidden layer after convergence of the training, the h being a higher-dimensional more-abstract feature vector of the x.
This invention relates to a computer program for facial recognition, specifically addressing the challenge of efficiently processing and compressing facial feature data for improved recognition accuracy. The program operates in an information processing device and performs several key steps. First, it collects facial images and performs uniform map processing to standardize the images. Next, the processed images are fed into a face recognition system to detect faces and extract facial feature data, which is then saved. A neural network model is set up to input this facial feature data. The model undergoes iterative training using an autoencoder-based method, where input data is compressed into a hidden layer, reconstructed, and compared to the original input to minimize prediction error. This process refines the model's ability to compress features accurately. After training, the model's parameters are saved. Finally, the trained model is used to obtain compressed feature vectors from new facial data, producing higher-dimensional, more abstract representations of the original features. This approach enhances feature compression efficiency while maintaining recognition accuracy.
2. The computer program product according to claim 1 , wherein a quantity of the facial images in step S 1 is more than ten thousand; each of the facial images has one and only one face; and for the map processing in step S 1 , a size of map is 1030p.
This invention relates to a computer program product for processing large-scale facial image datasets to generate a facial map. The technology addresses the challenge of efficiently organizing and analyzing extensive collections of facial images, particularly when each image contains only a single face. The system processes over ten thousand facial images, ensuring each image includes exactly one face, to create a high-resolution facial map with a size of 1030 pixels. The facial map is generated by mapping the facial images into a structured format, enabling efficient storage, retrieval, and analysis. The method involves extracting facial features from each image, normalizing the data, and compiling the results into a unified map. This approach improves accuracy in facial recognition tasks by reducing noise and ensuring consistency across the dataset. The invention is particularly useful in applications requiring large-scale facial analysis, such as surveillance, biometric identification, and demographic studies. By standardizing the input data and optimizing the mapping process, the system enhances computational efficiency and scalability.
3. The computer program product according to claim 1 , wherein the neural network model in step S 3 comprises a hidden layer having a node number being less than a node number of an input layer; and an input feature vector of the neural network model is the facial feature data.
This invention relates to a neural network model used for processing facial feature data. The model is designed to reduce computational complexity by using a hidden layer with fewer nodes than the input layer. The input to the neural network consists of facial feature data, which is transformed into a feature vector before being processed. The reduced node count in the hidden layer helps minimize computational overhead while maintaining accuracy in tasks such as facial recognition or analysis. The model is part of a broader system that extracts facial features from images or video frames, converts them into a structured format, and then applies the neural network for further processing. The architecture ensures efficient handling of high-dimensional facial data, making it suitable for real-time applications where computational efficiency is critical. The neural network's design balances performance and resource usage, making it adaptable for deployment in edge devices or cloud-based systems. The invention addresses the challenge of processing large-scale facial data efficiently without sacrificing accuracy, which is essential for applications in security, biometrics, and user authentication.
4. The computer program product according to claim 1 , wherein the parameter model from the x of the input layer to the h of the hidden layer of the trained neural network is only saved in step S 5 , and quantified by a Haisi quantization tool.
This invention relates to neural network model optimization, specifically quantizing trained neural network parameters to reduce computational and storage requirements. The problem addressed is the high resource consumption of neural networks due to large parameter sizes, which limits deployment in resource-constrained environments. The invention describes a method for training a neural network and selectively quantizing its parameters. During training, a neural network with an input layer (x) and a hidden layer (h) is trained using a training dataset. After training, only the parameters connecting the input layer (x) to the hidden layer (h) are saved and then quantized using a Haisi quantization tool. This selective quantization reduces the model size while preserving critical network functionality. The quantization process converts high-precision parameters into lower-precision representations, such as 8-bit integers, without significant accuracy loss. The remaining network parameters may remain in their original precision or be handled differently. This approach optimizes storage and computational efficiency, making the neural network more suitable for edge devices or real-time applications. The invention focuses on balancing performance and resource constraints by targeting specific parameter sets for quantization.
5. The computer program product according to claim 1 , wherein feature data x to be compressed in step S 6 is forwardly propagated from the input layer through the saved parameter model to obtain h which is the compressed feature.
This invention relates to a method for compressing feature data in a neural network model. The problem addressed is the computational and storage inefficiency of processing high-dimensional feature data in neural networks, particularly in scenarios where real-time processing or resource constraints are critical. The solution involves a technique for compressing feature data by forwardly propagating the data through a pre-trained neural network model to obtain a lower-dimensional compressed representation. The method begins by receiving feature data to be compressed. This data is then processed through a neural network model that has been previously trained and saved, with its parameters fixed. The feature data is propagated forward through the network's layers, starting from the input layer and passing through intermediate layers, until it reaches a designated layer where the compressed representation is extracted. The output of this layer is the compressed feature data, which is a lower-dimensional version of the original input. This compressed representation retains essential information from the original data while reducing its dimensionality, enabling more efficient storage and processing in subsequent applications. The technique is particularly useful in applications such as real-time data processing, edge computing, and systems with limited computational resources, where reducing the dimensionality of feature data can significantly improve performance and efficiency. The method leverages the pre-trained model's ability to extract meaningful features, ensuring that the compressed data remains informative for downstream tasks.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
April 16, 2020
February 15, 2022
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.